Some key technologies of neural network for structural damage identification, such as structural design of BP network, input parameter forms, training methods, learning rules and so on, are discussed. Modal parameters of a cylindrical shell structure under different support conditions are tested. Then these data are used as training samples for BP network, and depending on whether using random error, two different BP network is trained. Performance of these trained BP network in different case of test error is studied. The results show that adding a random error to the training samples can effectively improve the recognition ability of a BP network.